Glossary

This glossary defines recurring terms used across the wiki. It is intentionally concise; follow linked concept pages for formulas, examples, and production details. Use acronyms for short forms, notation for symbols, metrics for evaluation metrics, and references for source policy.

Term	Meaning
ALS	Alternating least squares, an optimization method often used for recommender matrix factorization.
ARIMA	Autoregressive integrated moving average, a univariate forecasting model.
Attention	A mechanism that weights different input positions when forming each output; the core operation behind transformers and attention.
Backpropagation	Reverse-mode differentiation that computes gradients of a loss with respect to network parameters.
Backtesting	Time-ordered evaluation that replays history to estimate forecast quality without leaking the future.
Bias-variance trade-off	The tension between underfitting (high bias) and overfitting (high variance) as model flexibility changes.
BM25	A lexical search scoring function based on term frequency, inverse document frequency, and document-length normalization.
Calibration	Agreement between predicted probabilities and observed frequencies.
Cold start	A recommendation or forecasting problem where a new user, item, or series has little history; see cold-start forecasting.
Concept drift	A change in the input-output relationship that degrades a deployed model over time.
Cross-entropy	A loss measuring the bits needed to encode outcomes from one distribution using another; standard for classification.
Data leakage	Information from outside the training fold that inflates offline scores and then fails in production.
Dimensionality reduction	Methods that compress features to fewer dimensions while preserving useful structure.
Embedding	A vector representation of text, images, users, items, or other objects; see embeddings.
Fine-tuning	Adapting a pretrained model to a task or domain by continuing training on new data.
Golden dataset	A curated, versioned evaluation set used for regression testing and model comparison.
Gradient descent	The iterative optimizer that steps parameters against the loss gradient.
Grounding	Conditioning generated text on retrieved or provided evidence so claims are supported.
Guardrails	Input and output checks that constrain a generative system to safe, valid behavior.
Hallucination	Fluent but unsupported model output; see hallucination mitigation.
JEPA	Joint Embedding Predictive Architecture, a family of representation-learning methods that predict in embedding space.
KL divergence	An asymmetric measure of how much one probability distribution differs from another.
MDP	Markov decision process, the formal model of sequential decision-making under reward.
Overfitting	Fitting noise in the training set so test performance drops; the high-variance end of the bias-variance trade-off.
Quantization	Reducing the numeric precision of weights or activations to cut memory and latency.
RAG	Retrieval-augmented generation, where retrieved evidence is used to condition a generative model.
Regularization	Penalties or constraints that reduce overfitting by discouraging overly complex models.
Reranking	Reordering an initial candidate list with a stronger, costlier model.
Stationarity	A time-series property where statistical behavior is stable over time; assumed by many classical models.
SVD	Singular value decomposition, a matrix decomposition into singular vectors and singular values.
Tokenization	Splitting text into the discrete units a model consumes.
Transformer	The attention-based architecture underlying most modern language and vision models.
Vector database	A store for embeddings that supports fast nearest-neighbor retrieval.
VLM	Vision-language model, a model that aligns visual inputs with language.

Application Domains

Application pages connect canonical methods to domain-specific inputs, outputs, risks, and evaluation slices. They carry the application topic in front matter so the wiki can surface them together.

Application	Canonical page	Main method families
Autonomous driving	Autonomous Driving	Perception, prediction, planning, control, simulation, safety evaluation.
Autonomous-driving evaluation	Autonomous Driving Model Evaluation	Scenario slicing, risk-weighted metrics, simulation, replay.
Demand prediction in logistics	Demand Prediction in Logistics	Forecasting, covariates, hierarchy, cold starts.
Energy forecasting	Energy Forecasting	Time-series forecasting, exogenous variables, calibration.
Malware classification and clustering	Malware Classification and Clustering	Classification, clustering, embeddings, adversarial drift.
Business message classification	Business Message Classification	Text classification, taxonomy design, evaluation.
Matchmaking	Matchmaking	Recommendation, ranking, constraints, fairness.
Medical MRI analysis	Medical MRI Analysis	Computer vision, segmentation, classification, patient-level validation.
Cultural heritage document extraction	Cultural Heritage Document Extraction and Entity Matching	OCR, entity matching, retrieval, human review.
News recommendation	News Recommendation	Recommenders, ranking, freshness, diversity, feedback loops.
Predictive maintenance	Predictive Maintenance	Time series, anomaly detection, survival/risk scoring.
Real-time action recognition	Real Time Action Recognition	Video understanding, streaming inference, trigger prediction.
Road scene perception	Road Scene Perception	Semantic segmentation, detection, pose estimation, tracking, scenario slices.
Gesture-based interaction	Gesture-Based Interaction	Gesture recognition, temporal localization, false-trigger control.

How to use this page

Use the glossary when a term appears before its full explanation. For study, jump from the term to the relevant subject area: matrix terms usually live in mathematical foundations, recommender terms in recommendation systems, retrieval terms in search, and governance terms in responsible AI.

Nav

Section — References and Glossary

Acronyms →